Diagnostic Performance of a Deep Learning Model Using Phonocardiogram and Single‐Lead ECG Signals for Detecting Moderate‐or‐Greater Aortic Stenosis: Multicenter External Validation

Figures

• 

  Figure 1. Flowchart of patient selection for external model validation. The initial validation cohort consisted of 973 patients. After excluding 22 patients because of missing PCG/ECG recordings (n = 13), lack of TTE–based assessment of AS severity (n = 9), or an interval longer than 6 months between PCG/ECG acquisition and TTE (n = 0), a total of 951 patients were included in the external validation cohort. Patients were subsequently categorized according to AS severity into moderate or greater AS (n = 70) and less than moderate AS (n = 881). AS, aortic stenosis; ECG, electrocardiogram; PCG, phonocardiogram; TTE, transthoracic echocardiography.

• 

  Figure 2. Overview of the Cardio‐EGG system for PCG and ECG acquisition and visualization. (A) Cardio‐EGG (AMI Inc., Japan) is a PCG device that enables acquisition of synchronized PCG and ECG signals within 8 s by direct placement on the precordial chest wall. (B) Real‐time visualization interface displaying ECG, heart sound waveforms, and time–frequency representations during signal acquisition. (C) Representative PCG waveform illustrating typical heart sound findings in a case of aortic stenosis. ECG, electrocardiogram; PCG, phonocardiogram.

• 

  Figure 3. Performance of the eAS model for detecting moderate‐or‐greater AS. ROC curves and corresponding performance metrics of the eAS model for detecting moderate‐or‐greater AS in the external validation cohort. (A) Model performance in all patients (n = 951). (B) Model performance in patients without prior valve replacement (n = 905). The AUROC with 95% CIs is shown in each panel. Sensitivity, specificity, PPV, NPV, positive LR, and negative LR are summarized below the ROC curves. AS, aortic stenosis; AUROC, area under the receiver operating characteristic curve; CI, confidence interval; LR, likelihood ratio; NPV, negative predictive value; PPV, positive predictive value; ROC, receiver operating characteristic.

• 

  Figure 4. Association between the eAS model output and key echocardiographic indices of AS severity. Sample sizes differ across panels because of variable‐specific missingness in the underlying echocardiographic measurements (Supporting Information S1: Table 4). (A) Distribution of echocardiographic parameters—peak aortic valve velocity (AV Vmax), mean transvalvular pressure gradient (AV mean PG), and AVA—according to the eAS model–defined groups. (B) Scatter plots showing the association between the eAS model output and corresponding echocardiographic measures of AS severity. Adjusted R 2 values are shown for each analysis. Data points are color‐coded according to echocardiographic AS severity. AS, aortic stenosis; AV, aortic valve; AVA, aortic valve area; PG, pressure gradient.

• 

  Figure 5. Subgroup analysis of the eAS model performance in the external validation cohort. Forest plots summarizing the performance of the eAS model across predefined clinical subgroups in the external validation cohort. For each subgroup, the AUROC, sensitivity, and specificity for detecting moderate‐or‐greater AS are shown with corresponding 95% CIs. The number of patients and the prevalence of moderate‐or‐greater AS are shown for each subgroup. Subgroups were defined according to age, sex, body mass index, AF, concomitant valvular heart disease, and LVEF. Missing data affected subgroup counts for BMI, AF, Moderate MR, and LVEF (Supporting Information S1: Table 4). AF, atrial fibrillation; AR, aortic regurgitation; AS, aortic stenosis; AUROC, area under the receiver operating characteristic curve; BMI, body mass index; LVEF, left ventricular ejection fraction; MR, mitral regurgitation; TR, tricuspid regurgitation.